325
AI Job interview warns against using AI yourself (thelemmy.club)
submitted 3 days ago by to c/mildlyinfuriating@lemmy.world
39 comments fedilink hide all child comments
 

Screenshot of a web-based AI job interview interface.

A pop-up titled "Quick note about AI usage" appears over the interview. It states that every response is checked for AI-generated content, that answers should be written from the applicant's own experience and in the applicant's own words, and that reviewers look at what a candidate did and how they did it rather than how polished the writing is. It adds that candidates who move to the next stage may be asked more about the examples given, so examples worth discussing should be chosen. A checkbox labeled "I'll answer in my own words." sits next to an OK button before the interview can continue.

Behind the dialog, the interview prompt asks for a specific example of working with others on a goal or project: what the person did to help and how everything was kept on plan, with the note that the example could come from home, work, school, a sporting team, or a community group. A tip at the bottom mentions that speech-to-text can be used to respond.

The juxtaposition of an AI-conducted interview warning against AI-generated answers is what the post highlights.

you are viewing a single comment's thread
view the rest of the comments
[–] 1 point 2 days ago (3 children)

That tester is old. It uses the technique of testing how predictable the text is, which is very inaccurate, especially on famous works that were in the training data. But recently, the major AI companies started watermarking their own outputs. It is now much easier to prove something is AI generated.

The watermarks are cryptographic in nature, and they rely on a secret key to both watermark and verify. So only the AI company can run the test, but it is accurate with long enough texts.

I'm sure its a cynical ploy to get people to upload future training data directly to the AI company's servers, but it does mean AI is more detectable now than it was.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] 2 points 2 days ago* (last edited 2 days ago) (2 children)

    Watermarks don't do shit in practical terms.

    a) simple text can't be watermarked.

    b) In any other media you can remove them by using ... another model! just run it through a local model - you dont even have to have an GPU for that task if speed isn't important, and any half-recent GPU can deal with any watermark in seconds in images, maybe minutes in audio, while videos take a bit longer depending on multiple factors.

    quick examples:

    AI is very good at recognizing even non-visible (for humans) watermarks and will remove them without question.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 1 point 2 days ago (1 child)

    The application process shown is clearly text based. Its not asking for images audio or video. The image and video tools you linked are designed for physical watermarks, not cryptographic ones. Its the same word, but very different in practice.

    The text watermarking is embedded into the output's word choice. It requires large amounts of text of text to be detectable, but it is there in plain text. Word choice is shifted by a biased random number generator based on a secret key. Without the secret key held only by the company, its not possible to tell which part the watermark is.

    Most things can be circumvented. There are local models that do not watermark. But most people are lazy, most people go the simple route of copying the output of a corporate AI product. People literally publish articles that end with, "is there anything more I can do to help."

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •